MACHINE LEARNING OPPORTUNITIES IN FREIGHT TRANSPORTATION OPERATIONS. NORTHWESTERN NUTC/CCIT October 26, Ted Gifford

Transcription

1 MACHINE LEARNING OPPORTUNITIES IN FREIGHT TRANSPORTATION OPERATIONS NORTHWESTERN NUTC/CCIT October 26, 2016 Ted Gifford

2 SCHNEIDER IS A TRANSPORTATION AND LOGISTICS LEADER WITH A BROAD PORTFOLIO OF SERVICES. 2

3 MACHINE LEARNING APPLICATIONS Pricing and Cost Estimation Spot & Contract Forecasting Capacity IC Drivers Driver Turnover Driver Safety Market Strength Forecasting TEXT ANALYTICS Capturing Lost Sales - Demand Forecasting Customer Relations Assessing Account Health 3

4 OPERATIONS CENTER SCHNEIDER NATIONAL 4

5 TEXT ANALYTICS TURNDOWN CAPTURE 10,000+ Tendered Orders/day Turndowns (Lost Sales) vary 5% to 20% Accurate Demand Forecasting requires understanding of total demand Difficult to capture via standard data entry CUSTOMER RELATIONS ~ 500 Customer Service Reps Gauge Customer Sentiment Improve CSR communication skills Pro-actively address issues that lead to churn or freight reduction Coaching and discover of systemic issues. 5

6 SOLUTION APPROACH (COMMON) 60K Messages/day , chat, phone Collect/ Store - Hadoop/MapR Preprocessing Parsing Regular Expressions Tokenizing NLTK (Python) Feature Extraction via TF-IDF Classification ensemble using Logistic Regression Stochastic Gradient Descent Multinomial Naïve Bayes Support Vector Machine Confidence score voting to determine message topic 6

7 DISTRIBUTION BY MESSAGE TOPIC 7

8 TURN DOWN ANALYTICS PROCESS FLOW 8

9 TURN DOWN CAPTURE SCHNEIDER NATIONAL Select messages associated with capacity request Use Content Extraction to create structured records Identify specific entities that define an order customer, geography, dates and times Uses Stanford NLP libraries as well as other pattern matching, and data exclusion techniques. Data is cleansed, normalized, matched to normalized locations, dates, and customers in Enterprise Master Data Checked against other sources to prevent duplication. Posted to Order System via Sqoop and ODI. 9

10 SENTIMENT ANALYSIS Use Sentence-based analysis Recursive Neural Tensor Networks (Socher et al.) Stanford NLP Library Deep Learning Sentiment Treebank includes sentiment labels for 215,154 phrases in parse trees of 11,855 sentences 10

11 SENTIMENT ANALYSIS Sentiment packages trained and built with social media and B2C communication do not work for B2B. In B2B communication expression of dissatisfaction or issues of concern is almost always emotionally neutral. Using a business-specific set of reference phrases, the Stanford Treebank is about 50% accurate. Underlying technology and algorithms (RNTN) is valid, but needs to be trained with business-specific and industryspecific phrases. Re-training with a set of 2200 business sentences, we have raised accuracy to 71%. Current plan is increased the training set until 80% accuracy is achieved. 11

12 SENTIMENT ANALYSIS SCHNEIDER NATIONAL Positive: We are good to go on this load!!! Negative: It s no longer acceptable to miss appointments without communication. This order is loaded, and we still haven t discovered why we are creating another service failure. Negative situation: Due to the weather advisory in the Gulf Region we are asking that you notify us of any potential weather delays and provide updates of managed loads that will be affected. The driver is updating a late ETD 8/ for delivery due to equipment related issues. 12

13 SENTIMENT TREND SCHNEIDER NATIONAL 13

14 THANK YOU